Drug Binding Affinity Predictions to Achieve Targeted Drug Discovery

新药的发现与对药物结合位点和整体药物-底物相互作用机制的详细了解密切相关。为此，我们一直在使用自由能映射的概念，它允许对药物与底物结合的所有可能方式进行侦察和加权。分子动力学技术的使用在这方面是至关重要的，因为它允许在热力学条件下进行模拟，这与实验/生理值相对应。由于活化过程的研究在分子动力学的标准方法中引入了内在的低效率，因此将使用先进的技术，包括元动力学，过渡路径采样和新实现的专有方法元射击。有效的分子动力学模拟依赖于有效的原子间力计算。对于生物系统，这是由经验力场保证的，因为相关系统的大小排除了使用任何从头算、密度泛函的技术，这些技术对于该领域的任何实际应用来说，计算量都太大了。在靶标系统中，我们重点关注DNA四聚体，即所谓的G4s，近年来由于其在癌细胞中的作用而被研究得更频繁。当G4s在DNA链的端粒末端形成时，端粒酶无法进入该位点以延长DNA序列。在癌细胞中，端粒酶通常会不间断地延长DNA序列，从而导致细胞永生。通过阻止这种永生化，癌细胞就不能无限复制，从而减缓肿瘤的生长。G4被金属离子稳定，这在之前使用元动力学研究金/NHC化合物的工作中也得到了证明。后者证明了分子动力学方法的有效性，但也指出了计算能量和力的方法的局限性。在本论文中，我们打算将先进的分子动力学技术与有机/生物系统的机器学习力场相结合，以达到更高的效率和精度。机器学习提供了一种更有效、更便携的方法来绘制势能面，通过这种方式，可以“学习”更高层次的理论，并随后在更便宜的计算时间尺度上表达出来，这使得以前非常不切实际甚至不可能的研究成为可能。机器学习电位的使用将允许以计算成本研究生物系统中的激活过程（键断裂，催化步骤，质子转移过程），可与力场相媲美。这将为研究生物系统中的药物作用以及活性步骤的能量学开辟全新的视角。除了对金属药物-底物相互作用和能量学的原始研究外，还将在所有力场模拟，QM/MM方法和机器学习电位之间进行基准测试，以比较和进一步优化后一种方法。

The discovery of new drugs is tightly connected to a detailed understanding of drug binding sites and overall drug-substrate interaction mechanisms. To this end we have been using the concept of free energy mapping, which allow for the reconnaissance and weighting of all possible ways a drug can bind to a substrate. The use of techniques of molecular dynamics is in this respect critical, as it allows for simulations at thermodynamic conditions, which corresponds to experimental/physiological values. Since the investigation of activated processes introduces an intrinsic inefficiency in standard methods of molecular dynamics, advanced techniques will be used, including metadynamics, transition path sampling, and the newly implemented, proprietary method metashooting.An efficient molecular dynamics simulation relies on an efficient calculation of interatomic forces. For biological systems this has been guaranteed by empirical force-fields, since the size of relevant systems excludes the use of any ab initio, density functional based techniques, which are simply computationally too heavy for any practical application in this area. Among the target systems, we are focusing on DNA tetramers, so called G4s, which have been studied more frequently in recent years due to their role in cancer cells. When G4s form at the telomere end of a strand of DNA, telomerase cannot access the site to extend the DNA sequence. In cancer cells the telomerase enzyme will normally extend the DNA sequence uninterrupted, which leads to cell immortalisation. By preventing this immortalisation cancer cells cannot replicate indefinitely, slowing the growth of a tumour. G4's are stabilised by metal ions, as also demonstrated in previous work focusing on Gold/NHC compounds using metadynamics. The latter has proved the efficiency of the molecular dynamics approach, but has also indicated limitations in the way energies and forces are calculated.In this thesis we intend to reach a superior level of efficiency and accuracy by combining technique of advanced molecular dynamics with machine-learned force fields for organic/biological systems. Machine learning is providing a more efficient, portable way to map potential energy surfaces, in a way that a higher level of theory can be "learned" and subsequently expressed on a computationally cheaper time scale, which allows for investigations previously very impractical or even impossible. The use of machine-learned potentials will allow to study activated processes in biological systems (bond breaking, catalytic steps, proton transfer processes) at computational costs, comparable to force fields. This will open completely new perspectives in the study of drug action in biological systems, and on the energetics of activated steps. Besides original investigations on metallodrug-substrate interactions and energetics, benchmarks will be performed, between all force-field simulations, QM/MM approaches and machine-learned potentials, for comparison and further optimisation of the latter method.